Anthropic Recruits OpenAI's Second Chip Engineer as $35 Billion Infrastructure Bet Reshapes the IPO Race

Clive Chan, who describes himself as the second hardware employee in OpenAI’s custom chip program, announced on June 6 that he has left OpenAI to join Anthropic. The move comes less than two months after Reuters reported that Anthropic was exploring custom chip design but had no dedicated team. Chan’s hire suggests that exploration has become recruitment.

The timing is not incidental. Anthropic filed its confidential S-1 at a $965 billion valuation. OpenAI is preparing its own public offering. Both companies are racing to demonstrate infrastructure independence to public market investors who will scrutinize their cost structures, supply chain dependencies, and margin trajectories. Custom silicon is the infrastructure story both are now telling.

What Chan Brings

Chan spent approximately 2.4 years at OpenAI building custom AI accelerators from scratch, according to Pulse2. He was involved in the strategic partnership between OpenAI and Broadcom, which targets chip deployment beginning in the second half of 2026. Before OpenAI, he spent two and a half years at Tesla’s Autopilot division working on custom training ASICs, handling software framework bring-up, datacenter co-design, and energy-efficient number formats.

In his LinkedIn announcement, Chan wrote that OpenAI’s chip team has “extraordinary” hardware talent density: “I don’t think there’s a better chip design team anywhere.” He described his motivation for leaving as wanting to “climb a new mountain from the bottom again,” according to The Decoder.

His new LinkedIn role description at Anthropic reads “perplexity per picojoule,” a phrase that could indicate either custom chip design (maximizing model performance per unit of energy in silicon) or software optimization for existing hardware. The ambiguity is likely intentional given the competitive sensitivity.

The $35 Billion Infrastructure Foundation

Chan arrives at a company that has already secured massive chip financing. Apollo Global Management and Blackstone finalized a $35 billion debt financing package to purchase Google TPU chips for Anthropic, Bloomberg reported via Investing.com. The transaction ranks among the largest private credit deals ever structured.

The deal uses a special-purpose vehicle that borrows funds, buys TPUs, and leases them to Anthropic. The structure keeps the obligation off Anthropic’s balance sheet entirely. Three tranches: $6 billion in A1 notes at investment-grade spreads (1 percentage point over Treasuries), $24 billion in A2 notes at a 5.75% coupon, and $4.5 billion in B notes at 8.5%. Broadcom backstops the senior tranches with a residual value support agreement, transferring credit risk from a high-growth startup to a $500 billion semiconductor company.

Broadcom CEO Hock Tan confirmed the arrangement, describing a broader “AI XPV platform” targeting more than 20 gigawatts of compute capacity through 2028.

Anthropic currently runs Claude on Google’s TPUs and Amazon’s Trainium chips. The Motley Fool reports that Anthropic’s compute expansion depends heavily on Amazon’s custom accelerators, with infrastructure suppliers like Coherent (whose data center segment posted $1.36 billion in Q3 FY2026 revenue) positioned as indirect beneficiaries of the IPO.

The Strategic Logic of Custom Silicon

The custom ASIC market is growing at 44.6% in 2026, three times faster than merchant GPUs, according to FourWeekMBA’s analysis of industry data. Every major AI infrastructure player has moved toward proprietary silicon: Google has TPUs, Amazon has Trainium, Microsoft has Maia, Meta is building custom inference chips, and OpenAI partnered with Broadcom.

For Anthropic, the business case centers on inference margins. Purpose-built silicon optimized for Claude’s specific architecture could deliver significantly lower cost-per-token at scale compared to general-purpose GPUs or even third-party custom chips. With inference costs becoming the dominant operating expense for frontier model companies, even a 20-30% efficiency gain translates into hundreds of millions in annual savings at Anthropic’s scale.

The OpenAI-Broadcom partnership offers a cautionary template. The Decoder previously reported that the partnership hit friction over production costs and OpenAI’s creditworthiness, with Broadcom reportedly unwilling to manufacture unless Microsoft committed to purchasing 40% of the output. Chan witnessed these negotiations firsthand. He now brings that institutional knowledge to Anthropic’s infrastructure strategy.

The IPO Infrastructure Arms Race

The timing of infrastructure moves relative to IPO filings is deliberate. Public market investors scrutinize AI companies on three dimensions: model capability, revenue growth, and infrastructure economics. Custom silicon addresses the third dimension by signaling that a company can control its cost structure as it scales.

Anthropic’s S-1 filing at $965 billion creates the equity incentive structure to recruit top hardware talent. Chan’s move suggests the pitch is working: join before the IPO, build from scratch, own the outcome through equity appreciation. FourWeekMBA notes that for OpenAI, “losing a chip lead to its most direct competitor is infrastructure knowledge walking across the street to the company filing for a trillion-dollar IPO.”

The competitive dynamic is recursive. OpenAI needs custom chips to demonstrate infrastructure independence from Microsoft (which has its own Maia chips and could theoretically deprioritize OpenAI’s needs). Anthropic needs custom chips to demonstrate independence from both Google (TPU supplier and investor) and Amazon (Trainium supplier and investor through AWS). Both companies need the infrastructure story to justify trillion-dollar public valuations.

From Exploration to Execution

In April 2026, Reuters reported that Anthropic’s chip plans were “in early stages” with “no dedicated team.” Two months later, the company hired a core member of the most advanced competing chip program. The progression from exploration to recruitment happened faster than most industry observers expected.

Chan indicated to Pulse2 that he cannot discuss details of OpenAI’s chip program, but pointed to previously announced deployment timelines beginning in H2 2026. That timeline means the first real-world performance data from OpenAI’s custom silicon will emerge just as Anthropic begins its own design process. Both companies will be watching each other’s infrastructure metrics as they pitch public market investors.

The Vertical Integration Endgame

The AI industry is consolidating around a pattern: frontier model companies that rent infrastructure eventually build their own. The question is no longer whether Anthropic will design chips, but how quickly Chan can assemble a team and produce a design competitive with what he helped build at OpenAI.

The $35 billion TPU deal ensures Anthropic has compute capacity for the next two to three years regardless of its chip program’s timeline. Custom silicon is the medium-term play: if Anthropic can bring its own chips online by 2028-2029, it would enter public markets with both a working product (Claude, currently generating revenue) and a credible path to infrastructure self-sufficiency.

For the broader market, the chip talent war between OpenAI and Anthropic signals that the AI infrastructure stack is vertically integrating at every layer. Model companies are becoming chip companies. Chip financiers are becoming AI infrastructure investors. The trillion-dollar IPO race is increasingly a story about who controls the silicon underneath the models.